The Direct and Indirect Influences of Interrelated Regional-Level Sociodemographic Factors on Heat-Attributable Mortality in Europe: Insights for Adaptation Strategies

Abstract

Background:

Heat is a significant cause of mortality, but impact patterns are heterogenous. Previous studies assessing such heterogeneity focused exclusively on risk rather than heat-attributable mortality burdens and assume predictors are independent.

Objectives:

We assessed how four interrelated regional-level sociodemographic predictors—education, life expectancy, the ratio of older to younger people (aging index), and relative income—influence heterogeneity in heat-attributable mortality burdens in Europe and then derived insights into adaptation strategies.

Methods:

We extracted four outcomes from a temperature–mortality study covering 16 European countries: the rate of increase in mortality risk at moderate and extreme temperatures (moderate and extreme slope, respectively), the minimum mortality temperature percentile (MMTP), and the underlying mortality rate. We used structural equation modeling with country-level random effects to quantify the direct and indirect influences of the predictors on the outcomes.

Results:

Higher levels of education were directly associated with lower heat-related mortality at moderate and extreme temperatures via lower slopes and higher MMTPs. A one standard deviation increase in education was associated with a $-0.46\pm 0.14$, $-0.41\pm 0.12$, and $0.41\pm 0.12$ standard deviation ($\pm \text{standard\hspace{0.17em}error}$) change in the moderate slope, extreme slope, and MMTP, respectively. However, education had mixed indirect influences via associations with life expectancy, the aging index, and relative income. Higher life expectancy had mixed relations with heat-related mortality, being associated with higher risk at moderate temperatures ($0.33\pm 0.11$ for the moderate slope; $-0.19\pm 0.097$ for the MMTP) but lower underlying mortality rates ($-0.72\pm 0.097$). A higher aging index was associated with higher burdens through higher risk at extreme temperatures ($0.13\pm 0.072$ for the extreme slope) and higher underlying mortality rates ($0.93\pm 0.055$). Relative income had relatively small, mixed influences.

Discussion:

Our novel approach provided insights into actions for reducing the health impacts of heat. First, the results show the interrelations between possible vulnerability-generating mechanisms and suggest future research directions. Second, the findings point to the need for a dual approach to adaptation, with actions that explicitly target heat exposure reduction and actions focused explicitly on the root causes of vulnerability. For the latter, the climate crisis may be leveraged to accelerate ongoing general public health programs. https://doi.org/10.1289/EHP11766

Introduction

Europe is warming, with a clear upward trend in average temperatures as well as an increasing frequency and intensity of heat waves.¹ When daily temperatures rise above an optimal temperature— known as the minimum mortality temperature (MMT) or percentile (MMTP)—the risk of heat-attributable death increases, with the rate of increase rising as temperatures approach extremes.² High temperatures have significant mortality impacts in Europe, with an estimated 108,000 heat-attributable deaths in people over 65 y of age in 2019 alone.³ At the same time, despite recent warming, the risk of heat-related mortality appears to be decreasing in some European countries,⁴ whereas in others it is stable or rising.¹ Further, the literature suggests there is considerable heterogeneity in mortality patterns across and within countries (although some apparent heterogeneity may be because of between-study methodological differences⁵).^1,6,7 Looking to the future, modeling suggests that, in the absence of climate mitigation and adaptation, heat-attributable mortality could rise substantially.^6,8,9 These current and potential health impacts necessitate adaptation¹; the heterogeneity suggests modifiable factors influence impacts, and this influence necessitates an understanding of the drivers of vulnerability to guide adaptation strategies.¹⁰

The contributors to heterogeneity in heat-related deaths have been investigated at the individual level (e.g., across people with different demographic, social and economic characteristics)^11–14 and at the area level (i.e., across locations).^15–20 Various social, economic, demographic, health-related, and material conditions have been found to influence risk.²¹ These include individual characteristics such as age, gender and health status, as well as contextual conditions such as climate, the built environment, green space, and the prevalence of air conditioning.¹ For systematic reviews of how such factors influence heat mortality, see Ellena et al.²¹ and Benmarhnia et al.¹⁰

In this paper we focus on sociodemographic factors (i.e., a mix of social and demographic factors). Overall, previous analyses have found that risk generally rises with age, the latter being represented by various variables (e.g., individual age, proportion of the population $>65$ y of age, life expectancy).^10,21 The risk implications of education- and income-related variables have been less consistent. For instance, in a cross-sectional comparative study of 340 cities in 22 countries (13 classed as developed countries and 9 as developing countries), Sera et al. found that the risk of heat-attributable mortality was higher at higher levels of education and gross domestic product per capita (GDPpc).²⁰ Chung et al. compared 47 prefectures in Japan, finding mortality risk in extreme heat was not strongly associated with income-related factors.¹⁹ Using individual-level data for the city of Turin, Italy, Ellena et al. found that higher education levels (in comparison with lower education levels) were associated with greater risk in men but lower risk in women.¹² Marí-Dell’Olmo et al. used individual-level data for the city of Barcelona, Spain, finding that although education tended to reduce risk, benefits did not rise monotonically with the level of education, and effects were less evident in women than men.¹³ Factors influencing the MMT in 400 cities (mostly from Southern Europe, North America, and East Asia) were investigated by Krummenauer et al.: Multivariate regression showed MMT was positively correlated with GDPpc.¹⁷

Notwithstanding methodological differences, previous analyses have used a common conceptual framing of the question of interest. That is, they investigated how the relative risk (RR) of heat-attributable mortality [or a related measure such as attributable fraction percent (AF%)] is modified by various factors, with each of the latter essentially being treated as independent. The findings, however, are potentially misleading. First, the heat-attributable mortality burden (i.e., the number of people who die due to heat) arises from a combination of the RR associated with a given temperature and the underlying mortality rate in the population (we define the latter as the average annual mortality rate per 10,000 in the population of interest). Specifically, RR is relative to the risk of mortality at the MMT, where this risk of mortality is itself a function of the underlying mortality rate. The same factors that influence relative risk are likely to influence (or be associated with) the underlying mortality rate; not accounting for this means the association between these factors and heat-related mortality may be over- or underestimated.

Second, many social, economic, and demographic factors do not exist independently: They change together and influence each other. For instance, it may be that an increase in education reduces risk via cognitive mechanisms and enhanced adaptive capacity²²; at the same time, education may contribute to population aging via social mechanisms (such as delayed and decreased population level fertility rates),²³ and a higher proportion of older people may increase both population level RR and the underlying mortality rate. That is, a factor that apparently reduces the heat-attributable mortality burden directly via reductions in RR may simultaneously increase the burden via indirect influences on RR and underlying mortality.

To address these gaps, this paper adopts a novel approach which allows us to consider the interrelations between a set of sociodemographic factors, the underlying mortality rate, and the change in RR across different temperature ranges. We focus on factors that capture major sociodemographic trends seen throughout Europe, i.e., rapid population aging and rising levels of educational attainment. These factors are readily quantifiable under the Shared Socioeconomic Pathways (SSPs); the development scenarios widely used in climate change projections.^24,25

Specifically, we use multilevel structural equation modeling (SEM)²⁶ to: a) quantify the direct relations between regional-level sociodemographic predictors and different aspects of the heat-mortality relation; b) investigate the indirect relations (i.e., via mediation) among the predictors and outcomes; and c) illustrate the importance of unspecified contextual factors. We then use the findings to derive insights into adaptation strategies. The predictors are: the proportion of people with upper secondary or tertiary education, life expectancy at birth, the proportion of older ($>64$ y of age) to younger ($<15$ y of age) people in the population and relative income (as the ratio of the average disposable income in a given region to the average disposable income in the country of which the region is a part, i.e., as within-country between-region income inequality). The four aspects of the heat–mortality relationship are: the average rate of increase in RR for each percentile rise in temperature at moderate temperatures (“moderate slope”); the average rate of increase of RR for each percentile rise in temperature at extreme temperatures (“extreme slope”); the MMTP; and the underlying mortality rate. The study is cross-sectional, focusing on 147 regions in 16 European countries, and it draws on the results of a previously published temperature-related mortality analysis (see “Methods” for further details and Figure S1).⁶ From here on, we use “Europe” as shorthand for the 16 European study countries.

Methods

In this section, we describe the study population and setting, the outcome data, and then the sociodemographic predictors and the control variables. Finally, we outline the model specification; that is, we describe the theory-based relations between the model variables in the form of an SEM.

Study Population and Setting

Our study draws its outcome data from a previous analysis of heat-attributable mortality in Europe,⁶ that focused on 147 regions in 16 European countries between 1998 and 2012 (Table 1; Figure S1). The regions were mostly NUTS 2 (“Nomenclature of Territorial Units for Statistics” at level 2), which are generally aligned with administrative regions in each country and are intended to have populations ranging from 800,000 to $3\text{\hspace{0.17em}million}$ (in practice, exceptions to this range exist, for instance, for geographical or cultural reasons).²⁷

Table 1.

Summary of the data used to fit the models as $\text{mean}\pm \text{SD}$, for 147 regions in 16 European countries over the period 1998–2012.

Geographic area and country	Number of observationsa	Heat-related mortality relative risk function			Underlying mortality rate per 10,000	Predictors
		Moderate slopeb	Extreme slopeb	MMTP (percentile)		Education (%)	Life expectancy at birth (y)	Aging index	Disposable income (Euros)
Central
Croatia	2	$1.14\pm 0.17$	$15.9\pm 0.3$	$83.7\pm 2.9$	$117.2\pm 9.5$	$77.4\pm 6.7$	$76.1\pm 1.5$	$1.4\pm 0.10$	—
Czech Republic	8	$0.47\pm 0.16$	$10.9\pm 4.3$	$87.1\pm 1.3$	$104.5\pm 2.1$	$90.5\pm 3.0$	$77.0\pm 1.0$	$1.2\pm 0.2$	$\mathrm{11,100}\pm \mathrm{1,569}$
Poland	16	$0.65\pm 0.16$	$15.3\pm 4.7$	$87.4\pm 1.5$	$96.7\pm 9.5$	$85.3\pm 2.8$	$75.4\pm 0.9$	$1.0\pm 0.14$	$\mathrm{10,106}\pm \mathrm{1,135}$
Slovenia	1	0.74	8.7	84.9	92.8	81.8	78.4	1.4	12,165
Northern
Denmark	1	0.32	4.2	87.1	104.1	74.2	78.4	0.98	14,225
Southern
Italy	21	$1.39\pm 0.46$	$44.1\pm 29.6$	$82.7\pm 1.5$	$100.8\pm 14.1$	$52.3\pm 6.2$ c	$81.7\pm 0.7$	$1.9\pm 0.44$	$\mathrm{16,014}\pm \mathrm{3,113}$
Portugal	5	$0.85\pm 0.096$	$41.8\pm 14.5$	$84.1\pm 1.4$	$109.2\pm 19.7$	$27.4\pm 7.4$	$79.2\pm 0.54$	$1.5\pm 0.37$	$\mathrm{12,480}\pm \mathrm{2,056}$
Spain	16	$1.00\pm 0.36$	$34.1\pm 8.5$	$83.0\pm 2.2$	$91.8\pm 12.9$	$50.4\pm 7.7$	$81.3\pm 0.8$	$1.6\pm 0.50$	$\mathrm{14,363}\pm \mathrm{2,501}$
Western
Austria	9	$0.71\pm 0.21$	$13.1\pm 6.5$	$84.9\pm 1.8$	$90.5\pm 13.2$	$79.8\pm 3.0$	$80.6\pm 0.7$	$1.3\pm 0.23$	$\mathrm{20,833}\pm 512$
Belgium	11	$0.62\pm 0.11$	$22.9\pm 6.4$	$86.0\pm 1.6$	$100.4\pm 11.7$	$68.7\pm 4.6$	$79.8\pm 1.3$	$1.2\pm 0.17$	$\mathrm{17,127}\pm \mathrm{1,749}$
France	22	$0.55\pm 0.16$	$17.7\pm 6.1$	$89.0\pm 2.1$	$93.7\pm 12.5$	$66.3\pm 6.1$	$81.1\pm 1.0$	$1.2\pm 0.26$	$\mathrm{16,900}\pm \mathrm{1,164}$
Germany	16	$0.56\pm 0.13$	$17.9\pm 6.2$	$83.9\pm 1.2$	$106.3\pm 8.5$	$85.5\pm 5.3$	$79.8\pm 0.7$	$1.9\pm 0.41$	$\mathrm{19,260}\pm \mathrm{1,914}$
Luxembourg	1	0.62	21.07	86.5	77.7	—	—	—	—
Netherlands	1	0.52	18.1	81.7	84.6	73.2	80.4	0.97	15,890
Switzerland	7	$0.60\pm 0.19$	$11.4\pm 4.24$	$86.2\pm 1.3$	$82.5\pm 5.6$	$78.8\pm 1.6$	$82.5\pm 0.6$	$1.1\pm 0.16$	—
England and Wales	10	$0.56\pm 0.14$	$17.1\pm 6.05$	$91.1\pm 0.96$	$98.3\pm 11.0$	$72.8\pm 3.1$	$79.9\pm 1.0$	$1.1\pm 0.16$	$\mathrm{16,527}\pm \mathrm{2,424}$
Europe	147	$0.76\pm 0.38$	$23.1\pm 16.86$	$85.8\pm 3.0$	$97.6\pm 13.0$	$69.4\pm 16.3$	$79.9\pm 2.2$	$1.4\pm 0.45$	$\mathrm{15,638}\pm \mathrm{3,615}$

Note: Education is the percentage of people with upper secondary or tertiary education; the aging index is proportion of older ($>65$ y) to younger ($>15$ y) people in the population. —, data not available; MMTP, minimum mortality temperature percentile; RR, relative risk.

^a “Number of observations” is the number of regions in each country.

^b Moderate and extreme slopes are change in RR of heat-related mortality for each percentile multiplied by 100, at moderate and extreme temperatures, respectively.

^c Two missing values.

Outcome Data: Four Aspects of the Heat–Mortality Relation

The four outcomes of interest—moderate slope, extreme slope, MMTP, and underlying mortality rate per 10,000—were derived from the above study.⁶ The methods used by Martínez-Solanas et al. are described in detail in the original paper⁶; the following provides a brief summary. For each region, a temperature–mortality exposure–response function (ERF) was estimated based on daily all-cause mortality data and daily mean temperature over the period 1998–2012, while controlling for day of the week, seasonal patterns, and long-term trends (i.e., mortality patterns differ by day of the week, vary with season, and change over time because of factors other than temperature; controlling for these allows the relation between temperature and mortality to be isolated).²⁸ The models were fitted using standard time-series quasi-Poisson regression in combination with a distributed lag nonlinear model (dlnm) in a two-stage process.^28,29 In stage one, region-specific temperature-mortality ERFs were fitted separately in each location; in stage two, between-region dependencies were exploited via a multilevel meta-analysis to derive the best linear unbiased predictions (BLUPs) of the regional functions.

From each of the above 147 temperature-mortality risk functions, we extracted the MMTP, as well as the RR at each percentile from the MMTP to the maximum temperature. We estimated the average rate of increase in RR for each percentile rise in temperature at moderate and extreme temperatures in each region using linear regression (outcome variable: RR; predictor: temperature percentile; regression coefficient for the predictor: rate of increase in risk for each percentile rise in temperature). Following convention,⁶ we used RRs from the MMTP (where $\text{RR}=1$ by definition) to the 97th percentile for moderate heat, and for extreme heat we used the RRs from the 98th percentile to the 100th percentile; we refer to the former as the “moderate slope” and the latter as the “extreme slope.” We defined the slopes as change in RR for each percentile rather than as change per degree of temperature because our study is comparative and thus focuses on a relative term; that is, our analysis compares how mortality risk rises with temperature, given the pattern of temperatures to which each population is accustomed.

Regional-level average annual mortality rate per 10,000 over the study period—i.e., the underlying mortality rate—was also extracted from Martínez-Solanas et al.⁶

The outcomes are interpreted as follows: The moderate and extreme slopes represent population vulnerability by quantifying how quickly mortality risk rises as temperature increases. We considered the moderate and extreme slopes separately because we expected sociodemographic factors would influence each of them differently. The overall risk across moderate temperatures arises from a combination of the MMTP (which defines the lower end of the range of moderate temperatures) and the moderate slope. For instance, for a given value of the moderate slope, a relatively high value for the MMTP would mean the overall risk across moderate temperatures is less than it would be if the MMTP had a relatively low value. The overall risk across extreme temperatures is represented by the extreme slope. The actual number of heat-attributable deaths in a population arises when RRs and a given temperature distribution are combined with the underlying mortality rate; that is, for a given pattern of RRs and a given temperature distribution, a lower underlying mortality rate would mean that the number of heat-attributable deaths were less than if the underlying mortality rate were higher.

Predictor Data: Four Sociodemographic Variables

We extracted regional-level data from the EURO-HEALTHY database (which was developed as part of a European project to provide a snapshot of the health of the European population over multiple dimensions and geographical levels),^30,31 which is a compilation of NUTS 2- and national-level health-relevant data derived from Eurostat (the statistical office of the European Union).³² We extracted the following variables: the proportion of 25- to 64-y-olds with upper secondary or tertiary education (“education”); life expectancy at birth (“life expectancy”); the ratio of people aged over 64 y to people under age 15 y (“aging index”); and, average disposable income (which is the balance of primary income and redistribution of income in cash, including social benefits³³). Rather than using the latter directly in our analysis, we used it to derive a measure of inequality between regions within each country. To do so, we divided the average disposable income of each region in a given country by the average disposable income of the country; i.e., within-country between-region inequality (“relative income”). We outline our reasoning for choosing these predictors in “Model Specification,” below.

We used data for the year 2007, which was the earliest year with nearly complete data for the regions of interest and falls roughly in the middle of the time period used to fit the temperature–mortality ERFs. Exceptions to the direct data extraction described were as follows: for disposable income, data for 2007 were sparse, so data for the year 2012 were used (2011 for Portugal); for countries where the initial temperature–mortality analysis was at the NUTS 1 (Germany, England, and Wales) or national (Denmark, Netherlands, Slovenia) level, NUTS 2-level data were aggregated using population weighting; for Switzerland, which was not included in EURO-HEALTHY, the predictors were extracted directly from Eurostat (and were thus equivalent to EURO-HEALTHY data).³²

Control Variables: Two Climate-Related Variables

To account for the influence of different climates on the region-specific ERFs, we followed standard practice and included controls for average annual temperature and the temperature range.³⁴ These data were available from Martínez-Solanas et al.⁶

Model Specification

We expressed the hypothesized relations between the model variables in the form of an SEM (Figure 1).²⁶ We discuss each of the predictors (shown in blue and labeled “A” in Figure 1) in turn, then outline the random and fixed effects (shown in orange and labeled “C,” and, in white and labeled “D,” in Figure 1, respectively) included in the model. Of note, although we expected the magnitudes and signs of the associations between the model variables to differ for each of the four outcomes, we expected the patterns of connection between the variables to be the same for each outcome; thus, Figure 1 shows the generic model used for all the outcomes.

Figure 1.

Model specification, with the predictors and the outcome of interest shown in blue (and labeled “A”), controls in gray (and labeled “B”), country-random effects in orange (and labeled “C”), and fixed effects for broad geographical regions (shown as “regional_indicator”) in white (and labeled “D”). Connections between the variables are shown as straight arrows. Unspecified causes that are external to the model, as well as measurement error, are represented by the disturbance terms (${\mathrm{\epsilon }}_1$ to ${\mathrm{\epsilon }}_4$), and disturbance correlations are indicated by the curved double headed arrows. Note: Education is the proportion of people with upper secondary or tertiary education; life expectancy is life expectancy at birth; the aging index is the proportion of older ($>65$ y) to younger ($>15$ y) people in the population; and relative income is the income of a region relative to the national average income.

The beneficial effects of education on health and mortality are well established, functioning via diverse processes associated with cognition, behavior, income, and social status.^35–38 The resulting combinations of ability to acquire information, capability to act in the face of hazards, everyday conditions of life, and health status suggest education is likely to reduce heat-attributable mortality, but that the effects may differ for different aspects of the heat–mortality relation. For instance, the capacity to take specific but temporary protective actions is likely to be of particular importance during extreme temperatures but is likely to be of limited relevance at moderate temperatures. Considering relations with the other predictors of interest, increased education has been shown to lead to increased life expectancy,^37,39 as well as higher incomes both at the individual and the aggregate level,^22,40,41 and lower income inequality.⁴² In addition, it is associated with lower fertility rates in the early stages of the demographic transition, leading to population aging later on.²³ Moreover, women with more education tend to delay child birth,^43,44 contributing further to population aging. In Figure 1, these postulated relations are shown as the arrows connecting education to the outcome of interest, relative income, life expectancy, and the aging index. The relations for the other predictors described below are shown in the figure in the same way.

Given that it is well established that older people tend to be more susceptible to heat-related mortality than younger people,^12,13,19,20 we included two predictors that capture different aspects of aging: life expectancy at birth and the aging index.

Life expectancy is frequently used as a summary indicator of population health.⁴⁵ However, it potentially has mixed effects on heat-related deaths. Older people are generally more susceptible to heat than younger people because of physiology⁴⁶; thus, population risk may rise with life expectancy. At the same time, a rise in life expectancy indicates the population has lower underlying mortality rates; this would be expected to reduce heat-related deaths. In addition, with rising life expectancy, poor health and reduced functioning (which increase the risk of heat-related mortality⁴⁶) may be “compressed” to ever older ages or may be “expanded” across old age⁴⁷; thus, susceptibility to heat-related mortality at older ages may fall or rise. Together, this suggests that higher life expectancies may contribute to both increasing and decreasing heat-related mortality, depending on which dimension of demography it represents.

The aging index captures age structure within a population. It would be expected that as the aging index rises, the risk of heat-related mortality would increase because, first, older people have greater susceptibility,⁴⁶ and second, as the proportion of older people rises, overall population risk would rise. In addition, the aging index may also capture mechanisms that reinforce or counteract this effect. Higher values of the aging index may indicate greater pressures on the health and social services⁴⁸; conversely, older societies may also be wealthier, perhaps rendering them both able and willing to adapt health and social services to an aging population; the potential correlation between the aging index and relative income is shown in Figure 1 by the double-headed curved arrow connecting them; see below.⁴⁹

We also considered whether life expectancy influences the aging index⁵⁰: As life expectancy rises there will be more older people, but the proportion of older people relative to younger people is also dependent on fertility rates and migration. We tested this relation using two alternative versions of our model. These took the same form as Figure 1 except the first included an additional direct pathway from life expectancy to the aging index (i.e., it assumed life expectancy had a direct influence on the aging index), whereas the second included the same additional pathway but also excluded the direct pathway from education to the aging index (i.e., it assumed education did not directly influence the aging index). Then, in each of the alternative models, we assessed the coefficient and standard error of the additional pathway. Because we found the link to be far from influential, we did not include it in Figure 1 (see “Results” for details). Of note, this empirical finding—i.e., that life expectancy and the aging index are not strongly associated—supports our theory-based decision to include both variables in the analysis.

Higher incomes and lower income inequalities have been linked to improved individual and population health in many studies.^37,51 Because Europe is currently characterized by converging national incomes but widening within-country income inequalities,^52,53 we analyzed the influence of the latter, represented as relative income (i.e., within-country between-region income inequality, as described above in “Methods” section “Predictor Data: Four Sociodemographic Variables”). We expect that the general health benefits of lower income inequalities^37,51,54 extend to heat-attributable mortality, such that regions in a country with a higher relative income have a lower average risk than regions with a lower relative income, with part of the effect being mediated through changes in life expectancy.³⁷ We also considered whether relative income may influence education. Although there is debate about the direction of the relation between education and income,³⁵ we assume that the dominant direction runs from education to relative income. That is, individuals with more education are likely to have higher incomes,^40,41 meaning that regions with a greater proportion of more highly educated people are likely to have higher average incomes.

Considered together, and as shown in Figure 1, the relations described above imply that—from the perspective of the model—education is an exogenous variable because it is not influenced by any of the other variables in the model; that is, it is free to vary without constraint from the other model variables. Life expectancy, the aging index, and relative income as well as the outcome of interest, are endogenous variables because they are influenced by other variables included in the model. Each of the endogenous variables also has unspecified causes that are external to the model; these are represented by the disturbance terms (${\mathrm{\epsilon }}_1$ to ${\mathrm{\epsilon }}_4$), which also capture measurement error. Furthermore, covariance (due to unspecified common causes) between endogenous variables that are not connected in the model is accounted for by disturbance correlations, shown by curved doubled-headed arrows (i.e., between the disturbances for relative income and aging and for life expectancy and aging). When fitting the model, we also tested for interactions (effect modification) between each pair of predictors; we assessed the interactions based on their coefficients and standard errors, considering them to be influential if the ratio of a coefficient to its standard error was roughly $<2$. For clarity, we do not show these in Figure 1 (We do show influential interactions in the fitted models; Figure 2.) In addition, we included the controls for regional climate (shown in gray and labeled “B” in Figure 1) for all outcomes except the underlying mortality rate (because the controls would be expected to have a limited influence on this).

Figure 2.

Model of best fit for each outcome, showing the path coefficients and their standard errors (in brackets) on the straight arrows, the error variances on the disturbance terms [circles containing $\mathrm{\epsilon }$, and the error covariances on the curved double headed arrows, as well as the total variance explained in the outcome (“var explained”)], for the moderate slope (A), the extreme slope (B), the MMTP (C), and the underlying mortality rate (D). Interaction terms, with their coefficients and standard errors (in brackets), are shown by the straight arrows that end on other straight arrows. For clarity, controls, random effects, and fixed effects are not shown. All outcomes and predictors are standardized. (Models cover 16 European countries over the period 1998–2012; number of regions in each models is: (A) $n=135$; (B) $n=144$, (C) $n=144$, (D) $n=144$). Note: Education is the proportion of people with upper secondary or tertiary education; life expectancy is life expectancy at birth; the aging index is the proportion of older ($>65$ y) to younger ($>15$ y) people in the population; and relative income is the income of a region relative to the national average income. Moderate and extreme slopes are change in RR of heat-related mortality for each percentile multiplied by 100, at moderate and extreme temperatures, respectively. MMTP, minimum mortality temperature percentile; RR, relative risk.

To account for unspecified country-level contextual effects on the outcomes of interest, we used random effects (shown in orange and labeled “C” in Figure 1). This approach uses partial-pooling across countries to estimate an intercept for each country (“random effect”), with the magnitude of the intercept being the expected mean value of the outcome of interest after accounting for the modeled variables.⁵⁵ That is, the random effects reflect between-country differences beyond those attributable to the variables of interest. For the upstream mediating relations, we used fixed effects for broad geographical areas (i.e., Northern, Central, Western, and Southern Europe) (shown in white and labeled “D” in Figure 1). The latter was partly a pragmatic decision because the reasonably small data set (i.e., 147 regions in 16 countries) did not permit the use of country-level random effects here, but there is also some theoretical grounding because countries in each broad area share some historical and political characteristics that may influence sociodemographic conditions (including health expenditure) and their influence on patterns of health.⁵⁶

We note that, for the three risk-related outcomes (i.e., moderate slope, extreme slope, and MMTP), the direction of all the arrows in Figure 1 is intended to correspond to the direction of the effects. For the fourth outcome— i.e., the underlying mortality rate—this correspondence also holds for the upstream connections. For the direct connections from the predictors to the mortality rate, however, the direction of the effect may run against the arrows. For instance, a fall in mortality leads to a rise in life expectancy rather than the other way around. Despite this, we use the same model specification for all four outcomes. This is because, from the perspective of our analysis, a question of central interest is: What is the association between the underlying mortality rate and a given set of sociodemographic factors that influences the risk of heat-attributable deaths? Given this question, we refer to relationships between variables either as effects or associations as appropriate.

In sum, the model represents the direct effects (or associations) between the predictors and the outcome of interest, the indirect effects (or associations) between the predictors and the outcomes via mediating variables, and unspecified country- and geographic area-level contextual factors (via random- and fixed-effects, respectively).

Fitting the Models

Before fitting the models, we standardized all the predictors and outcomes (i.e., standardized variables have a mean of 0 and a standard deviation of 1). We fit models for each of the four outcomes separately, beginning by including all the predictors and controls, then testing for interactions between each pair of predictors, and finally eliminating predictors that had little influence on the outcome (i.e., roughly, when the ratio of a coefficient to its standard error was $<2$). We then used likelihood ratio tests and AICs to compare our fitted model to equivalent null models with a) only the control variables (i.e., to test whether the predictors were contributing information) and b) without country-level random effects (i.e., to test whether using a multilevel model was justified).

All models were fit using the gsem command (“generalized structural equation modelling”) in Stata 16.1 (StataCorp).

After selecting the final models, we calculated the indirect (i.e., mediated) and total effects of the predictors on the outcomes using Statás “nlcom” (“Nonlinear combinations of estimators”) command.

Results

Data Description

Table 1 summarizes the outcome and predictor data, showing the study countries and the mean and variability [as standard deviation (SD)] for each variable. The data cover 16 European countries across four broad geographical areas, with an average of 9.2 observations (i.e., regions) per country (range: 1 to 21). For the moderate slope, the relative risk of mortality had a mean absolute increase of 0.0076 ($\text{SD}=0.0038$) for each percentile increase in temperature; the MMTP had a European average of 86% ($\text{SD}=3\%$). Combining these, the average range for moderate temperatures was from the 86th to the 97th percentile over which the average rise in RR was from 1 to 1.08. For extreme temperatures, RR had a mean absolute increase of 0.23 ($\text{SD}=0.17$) for each percentile rise in temperature. The average mortality rate per 10,000 across Europe was 97.6 with an SD of 13.

For the predictors, an average of 69% of people had upper secondary or tertiary education ($\text{SD}=16\%$) and life expectancy ranged from 75.4 y to 81.7 y ($\text{mean}=79.9$ y; $\text{SD}=2.2$ y). The average aging index was 1.4 ($\text{SD}=0.45$), and the average disposable income was 15,638 Euros ($\text{SD}=\mathrm{3,615}$ Euros).

Inspection of histograms for the outcomes showed the moderate and extreme slopes were skewed to the right, so we logged them when fitting the models; the MMTP and mortality rate were approximately normally distributed (Figure S2). Correlations between the four outcomes suggested they each captured different aspects of the heat–mortality relation (Table S1). Correlations between the predictors and controls suggested multicollinearity was unlikely to be a problem when fitting the models (Table S2) (In addition, in SEM, connected predictors are expected to be correlated).

Model Fitting

We found the model of best fit for each outcome by fitting models including all the predictors and controls (Table S3), then testing interactions between each pair of predictors (Table S4), and finally by eliminating predictors with little influence on the outcome (Figure 2; Table S5). We then compared the models of best fit to equivalent null models with a) controls only and b) without random effects. For all models, all the likelihood ratio tests and changes in the AIC suggested our selected models better predicted the outcome than the null models (Table S6).

As noted in the “Methods” section, we assessed whether life expectancy had an association with the aging index. We tested two alternative versions of the model: These corresponded to Figure 1, except a) with an additional arrow from life expectancy to the aging index and the disturbance correlation between the two variables removed and b) as for the previous modification but without the arrow from education to the aging index. In both cases the link was far from significant: The coefficients and standard errors were $-0.096$ ($\text{SE}=0.14$) and 0.13 ($\text{SE}=0.15$), respectively. Thus, we did not include this link in the models.

For the final models, the variance explained in the outcomes was 0.64 and 0.59 for the moderate and extreme slopes, respectively, and 0.80 and 0.81 for the MMTP and the underlying mortality rate, respectively (Figure 2). Scatter plots of model predictions vs. the observations used to fit the models show reasonably good fits (with correlation coefficients corresponding to the square roots of the reported variances explained); however, estimates tended to be biased upward at the lower end and downward at the upper end, which was most pronounced for the moderate and extreme slopes (Figure S3). Residual plots for the country random effects show the point estimates were approximately normally distributed (suggesting a reasonably good fit) (Figure S4).

Direct Associations between the Predictors and Outcomes

Table 2 shows the coefficients for the direct associations between the predictors (rows) and each of the outcomes (columns) in the final models. The values of the coefficients show how many standard deviations the outcome would be expected to change on average if the predictor increased by 1 SD.

Table 2.

Direct associations between the standardized predictors (rows) and the standardized outcomes (columns) in the final models, as point $\text{estimate}\pm \text{standard}$ error, for models covering 147 regions in 16 European countries over the period 1998–2012.

	log(Moderate slopea) ($n=135$)	Log(Extreme slopea) ($n=144$)	MMTP ($n=144$)	Underlying mortality rate ($n=144$)
Education	$-0.46\pm 0.14$	$-0.41\pm 0.12$	$0.41\pm 0.12$	—
Life expectancy at birth	$0.33\pm 0.11$	—	$-0.19\pm 0.097$	$-0.72\pm 0.097$
Aging index	—	$0.13\pm 0.072$	—	$0.93\pm 0.055$
Relative income	—	$0.12\pm 0.060$	—	—
Aging index X educationb	—	—	$-0.17\pm 0.048$	$-0.12\pm 0.046$
Aging index X relative incomeb	—	—	—	$0.15\pm 0.044$

Note: Coefficients are for direct effects of the predictors (rows) on the outcomes (columns), and show how many SDs the outcome would be expected to change on average for a 1 SD change in the predictor; the results are from multilevel structural equation models covering 16 European countries with controls from average annual temperature and temperature range (except in underlying mortality rate model), fixed effects for broad geographical regions, and country random effects. Education is the proportion of people with upper secondary or tertiary education; the aging index is the proportion of older ($>65$ y) to younger ($>15$ y) people in the population. —, no significant association; MMTP, minimum mortality temperature percentile; RR, relative risk; SD, standard deviation.

^a Moderate and extreme slopes are change in RR of heat-related mortality for each percentile multiplied by 100, at moderate and extreme temperatures, respectively.

^b “X” indicates an interaction between the two predictors.

A higher level of education was associated with lower heat-related mortality because of influences on all four outcomes. Risk was lower at moderate temperature via associations with the moderate slope ($-0.46$, $\text{SE}=0.14$) and the MMPT (0.41, $\text{SE}=0.12$), but the magnitude of the latter coefficient was reduced as the aging index rose ($-0.17$, $\text{SE}=0.048$). The extreme slope was also lower ($-0.41$, $\text{SE}=0.12$), as was underlying mortality, the latter via an interaction with the aging index (0.12, $\text{SE}=0.46$); these findings suggest that in populations with proportionally more older people, higher levels of education partly mitigate the rise in mortality.

Higher life expectancy had mixed effects on the heat mortality burden. Risk was higher at moderate temperatures via the moderate slope (0.33, $\text{SE}=0.11$) and the MMPT ($-0.19$, $\text{SE}=0.097$), but underlying mortality was lower ($-0.72$, $\text{SE}=0.097$). There was no association with the extreme slope. The aging index had no association with risk at moderate temperatures but was positively associated with both the extreme slope (0.13, $\text{SE}=0.072$) and the underlying mortality rate (0.93, $\text{SE}=0.055$). Higher relative income had a small positive association with heat-attributable deaths through the extreme slope (0.12, $\text{SE}=0.060$) and—via an interaction with the aging index—underlying mortality (0.15, $\text{SE}=0.044$).

Indirect and Total Associations between Predictors and Outcomes

In our models, the associations between education and the outcomes of interest (i.e., moderate slope, extreme slope, MMTP, and underlying mortality rate) are partly mediated by life expectancy, the aging index, and relative income (Figures 1 and 2). The direct, indirect, and total effects of education are shown in Table 3. The indirect effects of education on a given outcome are calculated by multiplying the coefficients along each indirect path from education to the outcome then summing the results; the total effects of education on a given outcome are calculated by summing the direct and indirect coefficients.

Table 3.

The direct, indirect, and total associations (rows) between standardized education and the four standardized outcomes (columns) in the final models, as point $\text{estimate}\pm \text{standard}$ error, for models covering 147 regions in 16 European countries over the period 1998–2012.

	log(Moderate slopea) ($n=135$)	Log(Extreme slopea) ($n=144$)	MMTP ($n=144$)	Underlying mortality rate ($n=144$)
Direct	$-0.46\pm 0.14$	$-0.41\pm 0.12$	$0.41\pm 0.12$	—
Indirect	$0.074\pm 0.036$	$0.16\pm 0.062$	$-0.15\pm 0.048$	$-0.16\pm 0.074$
Total	$-0.39\pm 0.14$	$-0.25\pm 0.13$	$0.26\pm 0.12$	$-0.16\pm 0.074$

Note: Coefficients are for direct, indirect and total effects (rows) of education (as the proportion of people with upper secondary or tertiary education) on the outcomes (columns) and show how many SDs the outcome would be expected to change on average for a 1 SD change in education; the results are from multilevel structural equation models covering 16 European countries with predictors for life expectancy at birth (not in the extreme slope model), proportion of older ($>65$ y) to younger ($>15$ y) people in the population (not in the moderate slope model), and relative income as the income of a region relative to the national average income (not in the moderate slope or MMTP models), controls from average annual temperature and temperature range (except in underlying mortality rate model), fixed effects for broad geographical regions, and country random effects. MMTP, minimum mortality temperature percentile; RR relative risk; SD, standard deviation.

^a Moderate and extreme slopes are change in RR of heat-related mortality for each percentile multiplied by 100, at moderate and extreme temperatures, respectively.

For the moderate slope, part of the influence of education was mediated via life expectancy, where the effects of education on life expectancy were both direct and indirect via relative income (Figure 2A). The coefficient for the total association between education and life expectancy (which includes a direct path and an indirect path via relative income) was $0.22\pm 0.077$ (as the point estimate and SE), and coefficient for the indirect association between education and the moderate slope via life expectancy was $0.074\pm 0.036$. This partially counteracted the direct effects, giving a total effect of $-0.39$ ($\text{SE}=0.14$).

For the extreme slope, education had indirect effects through relative income (0.069, $\text{SE}=0.039$) and the aging index (0.090, $\text{se}=0.051$) (Figure 2B), combining to give a total indirect effect of $0.159\pm 0.062$. These partly offset the direct effects of education on the extreme slope ($-0.41\pm 0.12$; Figure 2B), giving a total effect of $-0.25$ ($\text{SE}=0.13$).

For the MMTP, education had an influence through two indirect pathways (Figure 2C). The first was via life expectancy, which, as for the moderate slope, also involved relative income [$-0.043$ ($\text{SE}=0.026$)]. The second was via an influence on its own direct coefficient through its effect on the aging index ($-0.11$, $\text{SE}=0.038$). Combining these gives a total indirect effect of $-0.15\pm 0.048$. This offset the direct effects from the education to the MMTP ($0.41\pm 0.12$; Figure 2C), giving a total association of $0.26\pm 0.12$.

For the underlying mortality rate per 10,000, education had indirect associations via life expectancy (again involving relative income) ($-0.16$, $\text{SE}=0.060$) and via influences on the direct coefficient for the aging index (Figure 2D). For the latter, there were two paths. First, education is associated with relative income, which then interacts with the association between the aging index and the mortality rate; the coefficient was $0.085\pm 0.034$. Second, education is associated with the aging index, and education interacts with the association between the aging index and the mortality rate; the coefficient was $-0.081\pm 0.035$. When combined, these relations effectively cancel each other out; the net coefficient is $0.0048\pm 0.051$. Combining the indirect paths gives a total indirect association of $-0.16\pm 0.074$.

Overall Associations with the Burden of Heat-Attributable Deaths

For heat-attributable deaths at moderate temperatures, higher levels of education and higher life expectancy had opposite influences on risk (reducing and increasing risk, respectively) (Table 2; Figure 2A,C). The greater risk at higher life expectancies was countered, however, by the association between higher life expectancy and a lower underlying mortality rate, with the latter also being negatively associated with education (Table 2; Figure 2D). The influence of these two predictors was not independent: The beneficial effects of education on risk were partly offset by the influence of education on life expectancy, but this same relation contributed to higher reductions in underlying mortality (Table 3; Figure 2A–D).

For heat attributable deaths at extreme temperatures, higher levels of education were associated with lower extreme slopes, and higher relative incomes and aging indices were associated with higher extreme slopes (Table 2; Figure 2B). Further, higher levels of the aging index were associated with higher underlying mortality, with this relation strengthening at higher relative incomes (Table 2; Figure 2D). Again, the influences of the predictors were not independent: The beneficial effects of education were partly offset via its positive relations with the aging index and relative income, which led to higher risk and underlying mortality (Table 3; Figures 2B,D).

At both moderate and extreme temperatures, the burden of heat-attributable deaths is a function of both RR and the underlying mortality rate. Thus, even though the aging index is not associated with risk at moderate temperatures (Table 2; Figure 2A), the aging index still influences the number of deaths at moderate temperatures because of its positive association with underlying mortality (Table 2; Figure 2D). A similar situation holds for extreme temperatures, in this case for life expectancy, which does not influence risk but is negatively associated with mortality (Table 2; Figure 2D).

Unspecified Contextual Effects

For the direct effects of the predictors on the outcomes (i.e., moderate slope, extreme slope, MMTP, and underlying mortality rate), the country random effects represent modeled but unspecified country-level contextual effects. Table 4 shows these results as between-country variance and the 95% coverage interval for the random intercepts for each of the standardized outcomes.

Table 4.

Country random effects in the final models, shown as between-country variance and the 95% coverage interval of the random intercept (rows), for the four standardized outcomes (columns), for models covering 147 regions in 16 European countries over the period 1998–2012.

	log(Moderate slopea) ($n=135$)	log(Extreme slopea) ($n=144$)	MMTP ($n=144$)	Underlying mortality rate ($n=144$)
Between-country variance	0.25	0.061	0.44	0.21
Random intercept 95% coverage interval rangeb	$\pm 0.98$	$\pm 0.49$	$\pm 1.30$	$\pm 0.91$

Note: The results are from multilevel structural equation models covering 16 European countries with predictors for life expectancy at birth (not in the extreme slope model), proportion of older ($>65$ y) to younger ($>15$ y) people in the population (not in the moderate slope model), and relative income as the income of a region relative to the national average income (not in the moderate slope or MMTP models), controls from average annual temperature and temperature range (except in underlying mortality rate model), fixed effects for broad geographical regions, and country random effects. Education is the proportion of people with upper secondary or tertiary education. MMTP, minimum mortality temperature percentile; RR, relative risk.

^a Moderate and extreme slopes are changed in RR of heat-related mortality for each percentile multiplied by 100, at moderate and extreme temperatures, respectively.

^b This shows the range within which 95% of the country random intercepts would be expected to lie; for instance, for the log(moderate slope), 95% of the country random intercepts would be expected to have a value between $-0.98$ and $+0.98$ standard deviations of the log(moderate slope). Calculated as $\sqrt{between\text{-}countryvariance}\times 1.96$.

Unspecified contextual effects are smallest for the extreme slope, with a between-country variance of 0.061 and a 95% coverage interval for the random intercept of $-0.49$ to 0.49. In comparison, both the between-country variances and 95% coverage intervals are $\sim 4$ times larger for the moderate slope the underlying mortality rate (0.25 and $\pm 0.98$, and, $0.21\pm 0.91$, respectively), and take the largest values for the MMTP (0.44 and $\pm 1.3$).

Discussion

We assessed the influence of an interrelated set of regional-level sociodemographic factors on heat-attributable mortality across contiguous regions in 16 European countries. Previous studies have used individual- or aggregated-level data to assess how sociodemographic (and other relevant) factors directly influence RR and/or MMTP by analyzing single^12,13 or multiple cities^{14,16–18,20} or by comparing multiple subnational regions within a single country.^11,15,19 Our analysis adds to this body of work both conceptually and empirically.

Conceptually, we made three innovations. First, in contrast to previous analyses that assumed sociodemographic factors (and other predictors) are independent, we specified a model that accounts for between-predictor relations, thus assessing direct and indirect influences on heat mortality. Second, previous analyses focused on influences on relative risk; we additionally considered associations between the predictors and the underlying mortality rate in the population. RR at a given temperature refers to risk relative to the mortality rate at the MMTP, and this is in turn a function of underlying mortality rates. That is, the heat-attributable mortality burden arises from a combination of RRs and the underlying mortality rate (as well as temperature patterns). Consequently, focusing on RR alone may give misleading results; for instance, a predictor may have little influence on RR but may have a large effect on heat-attributable mortality through an influence on underlying mortality rates, or a predictor may have compounding or opposing influences on RR and underlying mortality rates.

Third, we explicitly separated the influences on risk at moderate (based on the MMTP and moderate slope) and extreme temperatures. Although extreme temperatures garner much attention, the majority of heat-related deaths occur at moderate temperatures.^13,34 Despite this, previous work has tended to focus on how various factors influence risk of all heat-related deaths in aggregate,²⁰ risks at extreme temperatures,¹⁹ and/or the MMT.^12,17

Empirically, we showed the different effects the sociodemographic factors have on each aspect of the heat-mortality relation; below, we provide a summary of the findings. Following this we discuss potential mechanisms underlying the observed relations and then some implications for adaptation strategies.

A higher proportion of people with an upper level of education was directly associated with a lower heat-related mortality burden via reduced risk at moderate and extreme temperatures as well as lower underlying mortality rates (Table 2; Figure 2). This makes intuitive sense and is broadly in line with previous work.^12,13 Alongside the beneficial direct associations, education had mixed indirect associations with the burden of heat-related deaths via positive associations with life expectancy, the aging index, and relative income (Table 3; Figure 2). For instance, education had a positive association with life expectancy, and, in turn, higher life expectancies were associated with higher risk at moderate temperatures; thus, education is associated with an increase in risk via this path.

Higher levels of life expectancy had mixed relations with the heat-related mortality burden, being associated with higher risk at moderate temperatures but lower underlying mortality rates (Table 2; Figure 2). In contrast, a higher aging index was consistently associated with a higher burden through higher risk at extreme temperatures and higher underlying mortality rates (Table 2; Figure 2). Previous analyses have assessed how age-related predictors influence risk^17,20 but have overlooked their opposing or reinforcing associations with heat-attributable deaths via underlying mortality.

Relative income made a smaller contribution to the heat-related mortality burden than the other predictors; it was associated with higher risk at both moderate (indirectly) and extreme (directly) temperatures but had a mixed association with underlying mortality (indirectly via higher life expectancy and by increasing the negative association with the aging index) (Table 2; Figure 2).

Our results for the influence of unspecified contextual effects (i.e., country random effects) showed these to be significantly smaller for the extreme slope in comparison with those for the moderate slope, MMTP, and the underlying mortality rate (Table 4). We suggest this makes intuitive sense. The risk at extreme temperatures would be expected to be more closely associated with innate human physiology than risk at moderate temperatures. This means that—given innate physiology is broadly similar across the study populations—after accounting for the study variables, the unexplained part of between-country differences in risk would be expected to be smaller for extreme temperatures in comparison with moderate temperatures. Or, put conversely, it would be expected that (after accounting for the study variables) unobserved modifiable factors would have a stronger influence on risk at moderate temperatures than at extreme temperatures. At the same time, however, such unobserved factors do influence the burden of heat-attributable deaths at extreme temperature via their apparent influence on underlying mortality rates.

We were not able to assess the mechanisms that link the predictors to the outcomes in our study, but some related inferences may be drawn from the results, including in relation to future research needs. For education, a range of mechanisms may be operating—involving, for example, knowledge, cognitive abilities, general health, and/or social status^35–38 —and the dominant mechanisms are likely to differ at moderate vs. extreme temperatures. During extreme heat, it is crucial that vulnerable people are aware of the risk, perceive themselves as at risk, and are able to take effective actions.¹ Thus, our results support recommendations that alerts should be targeted based on education levels not only in regard to ensuring knowledge but also in terms of viable actions of given groups.¹ In some contexts, it is likely that groups of lower social status may face restrictions that make them the least able to act. That is, mechanisms involving knowledge, cognition, and social status may all be crucial. For moderate heat, the most important effects of education may be via patterns of overall health status. Future work should aim to specifically investigate the key mechanisms, such has been done, for instance, using SEM in the more general case of the influence of education on health inequalities.⁵⁷

For the two factors associated with aging, we found each had different influences on heat-related mortality. Life expectancy and the aging index are similar in that higher values of both indicate a greater number of older people, and this would be expected to increase overall population risk. But, they also differ: Life expectancy is a general indicator of overall population health and standard of living,^45,54 whereas the aging index is an indicator of age structure.

For life expectancy, although older people may be inherently at higher risk of heat-related mortality than younger people, the size of the difference and whether the increase in risk begins in early or late old age is contingent on whether rises in life expectancy are accompanied by compression or expansion of poor health and reduced functioning.⁴⁷ Given that we found risk at moderate temperatures rises with life expectancy, and given that life expectancy is rising across Europe, we suggest that healthy aging,⁴⁹ which addresses the aging process over the entire life course, should be seen as an important part of adaptation to future climate change⁵⁸; we return briefly to this issue ahead.

The influence of the aging index on population risk during extreme heat may be simply because of a greater proportion of older people. The influence, however, may be strengthened by poor health and/or living conditions of older people. For instance, older people may be more likely to be socially isolated, live in poverty, or have mobility challenges: Such factors have been shown to increase risk.⁴⁶ The latter conditions are of course not inevitable features of older age, and addressing them would bring a range of benefits apart from reduced heat-related mortality. Further, a high aging index may be associated with more people living in elder care homes; given the wide variability of vulnerability among this group, it has been suggested that lowering indoor temperatures should be prioritized as a cooling strategy in these settings.⁵⁹

Finally, for relative income, the mechanisms driving the increase in risk at higher relative incomes at both moderate and extreme temperatures are unclear. It may be that higher relative incomes indicate a region is more urbanized (which may amplify exposure via the Urban Health Island effect⁶⁰) or that there are more income inequalities within the region (which may render populations more vulnerable⁵¹). This uncertainty reflects previous analyses, which collectively had mixed results for the influence of income-related variables.^17–20 Given that within-country inequalities are rising in Europe,^52,53 and the potential impacts of such inequalities on patterns of health and vulnerability,^37,51 future analyses—for instance, analyses that link income to area characteristics such as population density and green space—should investigate these relations further. In doing so, we suggest they should specifically consider between-predictor interrelations, as well as differences at moderate vs. extreme temperatures.

Our analysis has some potential implications for heat adaptation. Heat-Health Action Plans (HHAPs) are the cornerstone of protecting populations from present-day heat and adapting to climate change.^1,61 HHAPs incorporate a range of activities that include (but are not limited to) primary prevention (e.g., preheat event actions to reduce hazards and exposure)⁶² as well as Heat-Health Warning Systems (HHWSs), which combine weather forecasts and epidemiological studies to assess potential health impacts and communicate graded alerts.^61,62

By separating the influences of sociodemographic factors on the heat mortality burden into components attributable to, on the one hand, extreme temperatures, and on other hand, moderate temperatures and underlying mortality rates, our results point to issues relevant to each these HHAP activities. At extremes, heat is likely to play a major causative role in deaths (that is, heat alone may be a sufficient cause of death); our results for the unobserved contextual factors support this notion. Thus, HHWSs play a vital role in protecting health, and our results may help efforts to identify and target vulnerable groups. Of note, the recent World Health Organization (WHO) report on heat and health in Europe notes that although HHAPs usually refer to vulnerable groups, they often do not specify actions to address them.¹

In contrast, at moderate temperatures (at which most heat deaths occur), heat alone is unlikely to be sufficient to cause a death, with the ambient temperature often being only a contributing or precipitating factor. In addition, moderate temperatures are often not perceived as a threat by those at risk, which limits the potential role of a HHWS.⁵⁸ Further, at both moderate temperatures and extremes, the actual number of heat-attributable deaths is a function of the underling mortality rate, and, as our findings show, this is influenced by (albeit differently) the same factors that influence RR. Here, our results reinforce existing guidance that emphasizes the crucial role of generalized population health improvement in reducing the impacts of heat on health.^1,58

We do not intend to suggest the above distinctions are absolute: Some actions that reduce exposure during extremes will also reduce exposure to moderate temperatures, and addressing factors that influence susceptibility to moderate heat is also likely to have influences at extreme heat. Rather, we intend to raise the need to consider adaptation from dual perspectives: a “biophysical perspective” that tends to focus on technical, managerial, and behavioral interventions and a more “critical perspective” that attempts to address the root causes of vulnerability,⁶³ i.e., respectively, the need for actions that focus explicitly on responses to heat (e.g., climate-sensitive urban design⁶²) and actions that focus explicitly on processes that influence general population health. The latter are already acknowledged to be important, but despite this it has been found that there is relatively low uptake of the required long-term perspective and that HHAPs are often not a formal part of national health policies.¹ It is likely this is partly because of the limited resources available to HHAPs.¹ A possible way forward, then, is to integrate climate change considerations into already existing general public health programs, including those focused on healthy aging across the life course. That is, the climate crisis may be leveraged to give greater impetus to and to accelerate population health improvement. For instance, it could be emphasized that, for these health improvement programs to be successful, it is necessary to recognize that a warmer (and riskier) world will be an inevitable and constant presence in everyday life and that, when faced with this, a healthier population will be more resilient.

Finally, our results may be useful for projecting future heat-related mortality under various development (and climate) scenarios; that is, they could be used to model broad social adaptations. As described earlier (see “Introduction”), the sociodemographic factors chosen for the analysis are readily quantified in the SSPs. We believe such an approach to modeling would be a significant advance, because previous temperature–mortality projection studies have, for instance, used unchanging ERFs,⁶ arbitrary changes in the ERFs,^64,65 and changes in the ERF based on observed changes in other locations,⁶⁶ or they shifted the entire ERF to the right based on historical trends in the MMT.⁶⁷ Further, adopting an approach based on our analysis may allow more explicit assessments of the relative and combined benefits of heat-targeted vs. general population health-targeted actions.

Our paper has a number of limitations. First, we approximated the rise in risk at moderate and extreme temperatures using linearization. The original temperature-mortality ERFs were fit using sophisticated methods that allow for nonlinearities,⁶ but we sacrificed these precise risk estimates for two reasons: It facilitated the use of SEM, and our interest was not in risk itself but rather in underlying vulnerability as indicated by the average rate of rise in risk. Despite the loss in precision, we believe our findings provide useful insights into how sociodemographic factors directly and indirectly influence different aspects of the heat-related mortality burden.

Second, in the ERFs, heat was represented as daily mean temperature.⁶ Although this is a typical approach in time-series studies assessing temperature-related health impacts, other factors such as maximum temperature, minimum temperature, and humidity may also have some influence on risk.^61,68 Epidemiological studies, however, have generally found little evidence that temperature-based health impact predictions are improved when compound heat indices are used⁶⁸; thus, we suggest the use of mean temperature alone is sufficient for our purposes. In addition to daily changes in weather, the general climate type of a region may have some influence on our analysis: We accounted for this directly by including controls for average temperature and temperature range and indirectly through the country-level random effects. Although the latter does not account for between-region differences within a country, it seems reasonable to assume that any residual differences would be likely to have limited influence on how the sociodemographic factors influence heat-related mortality.

Third, we based our analysis on a single ERF for each region, and these were fitted using data covering 1998–2012 in a cross-sectional design.⁶ Consequently, we were unable to assess change over time, and this has implications in regard to changes that may have occurred over the last decade, as well as for assumptions about changes in the near future.

In terms of the previous decade, changes in weather and society may have contributed to changes in the relationships between our study variables and heat-related mortality. For changing weather, rising temperatures and more frequent exposure to extreme heat may have led to some physiological adaptation as well as greater awareness of the risks posed by extreme heat events^1,69; the latter, however, would not necessarily translate into adoption of protective actions.^1,58 For social change, many European countries have implemented Heat-Health Action Plans⁶¹ (although quantifying their health benefits have proven difficult⁷⁰) and changes in day-to-day living conditions that are not associated with our study variables may have either increased or decreased vulnerability in different groups and locations.^37,52 Given this, the nature of any shifts in the relations between the study variables and the outcomes are difficult to predict. We expect, however, that over the last decade the signs of the relations remain unchanged, and the magnitudes are likely to be broadly similar, partly because the mechanisms by which the variables influence heat mortality (and each other) are likely to have largely remained intact.

In terms of the near future, we also expect the relations between the predictor variables to remain broadly similar; this at least partially covers the window within which societies must adapt. In addition, because the vulnerability reduction-based adaptation strategies we discussed are very broad, they are likely to bring multiple health co-benefits regardless of any uncertainties around the future influences of sociodemographic conditions on heat mortality.

More generally, understanding how the influences of social variables on any health outcome may evolve over time presents challenges to most (if not all) statistically based analyses, including time-series study designs: The relations between sociodemographic variables and health outcomes are contingent across space and time. In fact, one of the purposes of vulnerability reduction is to actively change these relations such that risk associated with, for example, lower levels of education or aging is reduced. Thus, we argue that, although not without limitations, our findings provide useful insights but, as with all studies of this topic, should be considered in relation to the entire literature.

Fourth, as the ERFs were estimated at regional level,⁶ we used aggregated regional-level predictors, and this limits the potential to make causal inferences. In regard to this, there is a trade-off evident in previous studies: the data requirements of studies using individual-level predictor data (e.g., for education, age, or sex) usually restricts them to single cities^11–14; to compare multiple regions—which brings the benefit of greater variation in both the predictors and context—it is usually necessary to use aggregated data.^15–20 Although individual-level data is often seen as the “gold standard,”⁷¹ it may be argued that each approach asks different questions. Using individual-level data asks how individual qualities (e.g., years of education, age) influence average risk in a given population; using population-level data asks how population composition (e.g., percentage of individuals with higher education; percentage of older people) influence risk in a population, thus assessing the effect of the distribution of social resources across populations. That is to say, the limits of using population-level data are accompanied by gains, and this is also the case for individual-level data. Hence, our results would be complemented by—rather than replaced by—similar analyses using individual-level data.

Finally, we assessed just four sociodemographic predictors. This approach was intentional: The selected variables reflect major sociodemographic trends in Europe, and this methodology may facilitate health impact projections. However, we acknowledge that a wide range of evolving social, economic, and material (and other) factors influence heat-attributable mortality.²¹ Among these, the role of health services may be of particular interest because they may influence heat-related deaths at both moderate and extreme temperatures as well as via both risk and underlying mortality; we suggest that such an analysis could be facilitated by the use of health care typologies.⁷² More generally, although future work should include additional predictors, in doing so they should draw explicitly on empirically based theory to justify the inclusion of each variable and account for their interconnections.

Conclusions

Our findings show that when assessing how sociodemographic (or other) factors influence the heterogeneity in heat-attributable mortality, it is necessary (or at least beneficial) to consider how the factors influence each other, how they influence both risk and underlying mortality rates, and the differences in their influences at moderate and extreme temperatures. Further, our findings suggest that adapting to heat should involve not only actions to reduce exposure—such as those targeted at vulnerable groups during extreme temperatures—but also general public health actions that will reduce vulnerability and improve health across the entire life course.

The complexity of the temperature–health relation is widely recognized and accounted for using sophisticated statistical methods. Our results emphasize the need to also account for the complexity of how sociodemographic factors influence vulnerability and suggest that general progress toward healthier and more equitable societies will play a key role in protecting people from the impacts of climate change.